Privacy Policy
Effective date: October 23, 2026
Scope and operator
This policy describes the Prompt Monitor pilot extension and the publisher-operated management service at h-it.cloud. Organizations using another configured server administer their own deployment and must explain their own data practices to users.
What the extension reads
The extension reads text from the ChatGPT web composer when recognized text submission actions occur. The configured keyword list is checked locally. Full chat history, uploaded files, images, audio, ChatGPT replies and unrelated browsing history are not inspected by this pilot.
Optional DLP inspection and OpenAI processing
DLP is disabled by default. When an administrator enables DLP and approves processing, the extension transmits the current prompt text to the configured inspection server over HTTPS. Text may include passwords, authentication secrets, financial details, bank accounts, government identifiers, contact information, names, medical information, employee or customer records, confidential business information, code and legal material. The enabled categories determine enforcement; they do not restrict which content the submitted text may contain.
The server runs limited pattern checks. Locally detected blocking findings are returned without calling OpenAI. If GPT inspection is enabled and no such blocking finding is detected, the server sends the prompt after masking supported pattern matches to OpenAI's API for classification. Masking is incomplete: information not recognized by those patterns, including personal names or confidential material, can reach OpenAI in its original form. GPT receives only the current prompt and classification instructions, without the ChatGPT history. The model is asked to return category IDs only.
OpenAI API requests set store:false. This disables storage of the response as API application state; it does not guarantee zero retention, including abuse-monitoring retention. OpenAI processes data under its applicable API terms and data controls. Administrators must assess suitability and inform users before enabling inspection. A false negative can result in sensitive text being sent to ChatGPT.
Stored information
The extension stores the server origin, cached policy, last sync time and connection status locally. It does not persist prompt text. The backend does not intentionally store prompt text or model output in its database or application logs. Request data exists in memory while processing. Additional reverse proxies, tracing services or custom logs must be configured by the operator to avoid recording bodies.
Anonymous event reports contain only matched rule/category IDs and the action (warning, block or override). The backend adds a timestamp and event ID. No employee identifier is collected. Reports are unauthenticated and may be inaccurate or forged. A daily aggregate count of GPT request attempts is stored to enforce a request cap. The publisher's OpenAI API key is supplied only through the server environment; it is never included in the extension, policy response or event reports.
Public policy downloads include all keyword rule terms, labels, actions and category settings. Never place secrets in a wordlist. Policy changes, stored event reads, and DLP configuration require an administrator credential. Extension users do not need a token or registration.
Infrastructure logs
Apache access and error logs may contain IP addresses, request paths, timestamps, browser user-agent information, referrers and status codes. The standard access-log format does not record request bodies or Authorization headers. Hosting providers may process infrastructure data. The extension does not collect GPS information.
Purpose, sharing and retention
Information is used to check prompts for configured sensitive categories, apply warning or blocking rules, synchronize policies, report anonymous policy events and operate the service. Prompt text is shared with the configured inspection server and, when GPT is enabled, with OpenAI as described above. Data is not sold or used for advertising, unrelated purposes, creditworthiness or lending decisions.
Local settings remain until cleared or the extension is removed. New event ingestion prunes records older than 30 days and retains at most 10,000 events; cleanup is not scheduled when no reports arrive. Daily request counts are pruned to the current UTC date when a GPT request is reserved. Rules and API credentials remain until changed or removed by the operator. Infrastructure logs and backups have separate operator-controlled retention. Old registration records from earlier pilots may remain in historical databases or backups.
Choices and limitations
Administrators can disable DLP, disable GPT, or choose off, warn or block per category. Semantic categories require GPT. Users can cancel or explicitly approve warnings; blocks have no override. Inspection failures, request-cap exhaustion and timeouts stop recognized send actions. Users can disable or remove the extension subject to organization device policies. This pilot is a UI guard, not a complete network DLP boundary: it cannot prevent all leaks, cover unsupported interfaces, or prevent ChatGPT's own page code from processing composer text independently. Enabling GPT also introduces API cost. The public inspection endpoint has a shared request cap and limited concurrent scans; it is a pilot and is vulnerable to quota exhaustion by unauthenticated clients.
Contact
Questions and deletion requests: lior.harel291@gmail.com. For another organization's deployment, contact its administrator.